OpenAI's Quiet Play: ChatGPT Work, Codex, and the Ona Acquisition

GPT-5.6's July 9 rollout got the headlines. Three tiers, Sol at the top. And the model is the least interesting thing OpenAI shipped this cycle.

The interesting part is the stack they quietly built around it. Three pieces. ChatGPT Work is an agent that researches across your connected apps and files, then hands back finished documents, spreadsheets, presentations, and reports. Not answers. Deliverables. Codex is now past 5 million weekly users. And the Ona acquisition, the German startup formerly known as Gitpod, gives Codex persistent cloud-based operation. Agents that keep working after you shut the laptop.

The pattern

Every piece points at the same target. Knowledge work that ends in an artifact. The industry spent two years selling chat, and chat doesn't show up in a P&L. A finished quarterly report does. A merged pull request does.

We think about this constantly, because it's the lesson we relearn on every client build. Automation only counts when it completes the deliverable. An AI that drafts 80% of an invoice has created a review job. An AI that files the invoice, correctly, with an audit trail, is the one that takes hours out of your week.

What to do about it

If your team lives in the OpenAI ecosystem, ChatGPT Work is worth a pilot on one contained workflow. Weekly reporting is the obvious candidate. Watch two numbers: hours actually saved, and how often a human had to fix the output. If that second number is high, congratulations, the tool invented a new job instead of removing one.

And if you'd rather someone else ran that experiment and handed you the wiring that already works, well. Hello. That's the whole job.